基于多粒度增强和答案验证的法律文书阅读理解模型  

Legal text reading comprehension model based onmulti-granularity enhancement and answer verification

作　　者：王婧蕾 徐建[1] Wang Jinglei;Xu Jian(School of Computer Science&Engineering,Nanjing University of Science&Technology,Nanjing 210094,China)

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机应用研究》2025年第3期700-707,共8页Application Research of Computers

摘　　要：近年来法律文书阅读理解逐渐成为一个研究热点,它要求模型能够利用有限的数据完成答案分类以及证据和答案的提取,然而现有模型编码粒度单一,且问题和法律文书之间仍缺乏充分的交互。为了解决上述问题,提出了一种基于多粒度增强和答案验证的法律文书阅读理解模型。该模型首先使用多粒度编码模块,以字符、词语和序列三种粒度来捕捉输入文本中多层次的信息,提高模型提取信息的能力;然后引入基于问题和证据的注意力,增强问题和法律文书之间的交互,深入挖掘答案分类的线索信息;最后受到人类阅读过程的启发,该模型采用答案验证机制结合局部和全局信息来预测答案。在中文法律阅读理解数据集CAIL2019、CAIL2020和CAIL2020-Enhanced以及英文数据集HotpotQA上的实验结果表明,所提方法的joint F 1分别为76.48%、64.16%、70.82%和69.39%,优于基线模型。In recent years,legal text reading comprehension has gradually become a research hotspot,requiring models to complete answer classification and extraction of evidence and answers with limited data.However,existing models encode only at a single granularity and still lack sufficient interaction between questions and legal texts.In order to solve these problems,this paper proposed a legal text reading comprehension model based on multi-granularity enhancement and answer verification.The model firstly used the multi-granularity encoding module to capture multi-level information from the input texts at three granularities:character,word and sequence,which improved the model’s ability to extract information.Then,it introduced an attention based on questions and evidence to enhance the interaction between questions and legal texts and to deeply mine the clues for answer classification.Finally,it proposed an answer verification mechanism inspired by the human reading process to predict answers by integrating local and global information.Experimental results on the Chinese legal reading comprehension datasets CAIL2019,CAIL2020,CAIL2020-Enhanced and the English dataset HotpotQA show that the proposed method achieves 76.48%,64.16%,70.82%and 69.39%joint F 1 respectively,outperforming the baseline models.

关 键 词：法律文书阅读理解 多粒度编码模块 注意力机制 答案验证机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]